Pith. sign in

REVIEW 4 major objections 6 minor 21 references

AI-Based Thermal Video Analysis in Privacy-Preserving Healthcare: A Case Study on Detecting Time of Birth

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read Thermal video plus a 3D CNN estimates time of birth to within about 1.5 seconds.

desk verdict A genuinely new privacy-preserving clinical video application, but the headline accuracy doesn't match its own table and the evaluation protocol needs tightening. read the letter →

arxiv 2502.04365 v1 pith:RUUIZWFV submitted 2025-02-05 cs.CV cs.AI

classification cs.CVcs.AI
keywords timeofbirthdetectionthermalvideoanalysisprivacy-preservinghealthcarenewbornresuscitationspatiotemporalclassificationMoViNetGaussianmixturemodelnormalizationvideo-basedclinicaldocumentation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to show that the time of birth can be detected automatically and with second-level precision from thermal video, without capturing identifiable images of mothers or healthcare staff. The proposed system classifies short sliding-window clips of ceiling-mounted thermal footage as birth or no-birth, smooths the per-second scores, and reads out the first moment the smoothed score crosses a confidence threshold. On a test set of 25 deliveries it detects the birth in 96% of videos, with a median absolute error of 1.5 seconds against manual annotation (the abstract states 1 second). This matters because time of birth is currently logged by hand, usually to the nearest minute, and the first 60 seconds after birth are the window in which resuscitation decisions are made.

What carries the argument

The load-bearing object is the sliding-window classifier over GMM-normalized thermal video. Each second, the model sees the previous 25 frames and outputs a birth probability; the sequence of probabilities is smoothed by a 3-tap rectangular FIR filter, a simple moving average over the last three prediction scores, and the estimated time of birth is the first timestamp at which the filtered score reaches a fixed threshold of 0.9. The GMM normalization fits three Gaussian components to each video's temperature distribution (background, clothing and bedding, and skin), rescaling intensities around the skin-temperature peak so that predictions depend on relative heat rather than absolute camera readings.

What would settle it

Have two or more clinicians independently re-annotate the 25 test videos (and ideally a fresh set of births) without seeing the model output, and measure inter-annotator agreement. If the spread between human annotators is comparable to or larger than the reported 1.5-second median error, the system's accuracy against a true time of birth is not established; if the model matches human consensus better than individual humans match each other, the claim is strengthened.

Watch

Extended reading notes

Core claim

The paper's central claim is that spatiotemporal information from thermal video is sufficient to pinpoint the moment of birth, and that a lightweight 3D CNN (MoViNet-A2) with a 3-sample smoothing filter and a threshold of 0.9 achieves clip-level precision of 91.4% and recall of 97.4%, estimating ToB in 24 of 25 test videos with a median absolute deviation of 1.5 seconds (abstract: 1 second). The method outperforms the authors' earlier image-based approach, whose median error was 3.2 seconds and mean error 87.8 seconds under the same threshold, by exploiting temporal birth dynamics instead of single frames. The system also demonstrates that relative temperature normalization via a Gaussian mixture model can make the network robust to camera autocalibration and room-temperature variation while preserving privacy.

Load-bearing premise

The evaluation assumes that the manual ToB annotations, produced by carefully watching the same thermal videos, are accurate enough to serve as ground truth; if those human timestamps are biased or noisy, the reported precision measures agreement between the model and one human reading of the same camera feed rather than true birth time.

Editorial extensions

If this is right

  • Time-of-birth documentation can move from manual, minute-precision logging to automatic, second-precision timestamps generated from ceiling cameras.
  • Resuscitation debriefs and research timelines can be built from objective video-derived timestamps, including the interval from birth to first stimulation or ventilation.
  • The same privacy-preserving thermal pipeline may be reused for other clinical events that have a clear thermal or movement signature, such as cord clamping or first breath.
  • The threshold and filter settings give a tunable trade-off: raising the threshold suppresses false detections but increases the chance of missing a birth in positions where the newborn is hard to see.
  • A fallback image-based detector can cover the limited-visibility cases where the video-based model fails, as the paper explicitly proposes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If thermal-video timestamps are ever used in clinical documentation, the ground-truth question shifts: the current evaluation compares the model with a human reading of the same thermal feed, so the reported accuracy is agreement between two readers of one modality, not validation against an independent definition of birth time.
  • The non-causal whole-video GMM normalization means the reported numbers are for retrospective analysis; a causal, streaming version of the normalization would be needed before the system can issue real-time alerts during a delivery.
  • Because the test set is small (25 videos), excludes twins, and includes a known failure in the hands-and-knees position, the second-level precision claim should be re-measured on a larger, more diverse cohort before deployment decisions.
  • The same relative-temperature reasoning could transfer to other privacy-sensitive settings where timing matters, for example documenting when an intervention starts or when a procedure ends in emergency care.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a spatiotemporal deep-learning system to detect the Time of Birth (ToB) from thermal videos in a privacy-preserving manner. Videos are normalized with a GMM-based method, fed as sliding-window clips to a 3D CNN (I3D, X3D, MoViNet) trained for binary ToB/no-birth classification, and the predicted scores are post-processed with an FIR filter and thresholded to estimate ToB. On a test set of 25 manually selected thermal birth videos, the MoViNet-A2 backbone achieves 91.4% precision and 97.4% recall at clip level, and the inference pipeline yields a median absolute error of 1.5 s (Q2=1.5 s) with ToB found in 96% of videos, while the abstract and conclusion claim a 1-second median absolute deviation. The authors compare against their earlier image-based method.

Significance. If the reported accuracy holds, this would be a valuable step toward automated, second-level ToB documentation during newborn resuscitation, with the added privacy advantage of thermal imaging. The paper includes a reasonable comparison of multiple backbones, a causal inference design with FIR filtering, and a clear statement of the clinical motivation. However, the strength of the quantitative claims is limited by the fact that the threshold and model are selected using the same test set, and the ground-truth labels are derived from the same thermal modality; the headline numbers should therefore be interpreted as model-human agreement under favorable selection, not as a validated clinical measurement. The work is nonetheless a useful case study and the pipeline is well described.

major comments (4)
  1. [Abstract, Section 5 (Table 2), Section 6] The abstract and conclusion state that the system estimates ToB with an 'absolute median deviation of 1 second' from manual annotations, but Table 2 reports Q2 = 1.5 seconds for the final video-based method with FIR filter, and the conclusion repeats the 1-second claim. This internal inconsistency directly misstates the central quantitative result and must be corrected, with the abstract and conclusion updated to match the actual Q2 value or with an explanation of why 1 second is reported.
  2. [Section 4.2, Section 5 (Figure 4a, Table 2)] The confidence threshold gamma = 0.9 is chosen after inspecting the FPR curve (Figure 4a) computed on the same 25-video test set, and the same set is then used to report the ToB errors in Table 2 and Figure 4b. This constitutes selection on the test set and makes the reported Q2 and B.F. values optimistic. A separate validation set (or nested cross-validation) should be used for threshold selection, and if that is not possible, the paper should explicitly state the bias and provide a sensitivity analysis over thresholds on a held-out set.
  3. [Section 2, Section 4.2, Eq. (6)] The manual ToB annotations used as ground truth in Eq. (6) are obtained by carefully inspecting the same thermal videos (Section 2). Consequently, the reported median absolute error of 1.5 s (Table 2) measures agreement between the model and a human reading of the same thermal feed, not accuracy against an independent clinical definition of birth time (e.g., the head, torso, and nates fully outside the perineum). This premise propagates to every error statistic in Exp. 2 and to the clip-level labels. Please discuss this limitation explicitly, describe the annotation procedure in detail, and, if feasible, validate the annotations against an independent source (e.g., obstetric records, RGB video, or multiple annotators).
  4. [Section 4, Table 1, Section 5] The model selection in Exp. 1 is performed on the test set: Table 1 compares backbones on the same 25-video test set, and the best-performing model (MoViNet-A2) is then used for the final evaluation. This double use of the test set for model selection and performance reporting compromises the statistical validity of the reported precision, recall, and MCC values. With 321 videos, a dedicated validation split (or repeated cross-validation) should be feasible; at minimum, the paper should acknowledge this limitation and provide estimates of the selection bias.
minor comments (6)
  1. [Section 4.1.2] The sentence 'Weight decay of 0.97 is applied every 1k steps' appears to mix learning-rate scheduling with weight decay; please clarify whether this is a learning-rate decay factor or an L2 regularization weight decay schedule.
  2. [Section 4.1.2] The classification threshold used to compute precision/recall in Exp. 1 is not specified; please state whether a default 0.5 threshold or a tuned threshold was used.
  3. [Section 5, Table 2] Table 2 reports quartiles and B.F. for the video-based method, but it is not explicitly stated that Q1/Q2/Q3/mean are computed over the 24 videos where ToB was found (excluding the missing video); please add a note to that effect.
  4. [Section 5] The sentence 'It is important to note that the dataset used in this work is larger than the one used in the image-based method' is potentially confusing because Table 2 is said to use the same 25 test videos; clarify whether the image-based model was retrained on the enlarged dataset or evaluated as originally trained.
  5. [Section 2] The test set is described as 'manually selected,' and it is not stated whether the selection criteria (other than excluding twins and preserving position distribution) could introduce bias; please describe the selection process and its potential impact.
  6. [Section 3.1] The GMM normalization uses 'empirical constraints' for the temperature range of interest; please provide the specific constraint values so that the method is reproducible.

Circularity Check

2 steps flagged · score 6.0 of 10

Exp. 2's headline ToB accuracy is partly fitted: the threshold gamma=0.9 is selected from the test-set FPR curve before reporting the 96%/1.5 s statistics, and the ground truth is a human reading of the same thermal videos.

  1. fitted input called prediction [Section 3.3 (Eq. 4) and Section 5 (Fig. 4a, Table 2)]
    "To estimate the ToB, we identify the first timestamp where the filtered scores exceed a predetermined confidence threshold γ. ... Figure 4a depicts the False Positive Rate (FPR) of yh(t) at various thresholds ... By setting a threshold of γ = 0.9, we examine the time difference between our predicted ToB and the manual annotations across the test set."

    Eq. 4 defines the ToB estimate as the first threshold crossing of the filtered score. The threshold gamma is not fixed a priori: the paper plots FPR on the same 25 test videos and then chooses gamma = 0.9. Table 2's Exp. 2 statistics (96% found, median |err| = 1.5 s, mean 2.1 s) are computed on those same videos after that choice. Hence the headline 'prediction' is conditional on a parameter fitted to the test set, so the reported detection rate and error are partly fitted values rather than out-of-sample predictions.

  2. other [Section 2 (label generation), Section 4.2 (Eq. 6), Abstract]
    "The registration of ToB in Liveborn was sometimes delayed due to all the activities that happened around the birth. Therefore, in this study, manual annotation of the ToB with second precision was done by carefully inspecting the thermal videos. ... For evaluation purposes, we define the error err as the time difference between Tbirth_hat and the manual annotated ToB (Tbirth): err = Tbirth_hat - Tbirth (6)"

    The reference ToB in Eq. 6 is operationalized as a human timestamp produced by inspecting the same thermal video stream the model ingests, not by an independent clinical measurement (e.g., obstetric record or RGB observation). The abstract's 'absolute median deviation of 1 second compared to manual annotations' is therefore an agreement statistic between two readers of the same modality, model and human, rather than a validation against an external definition of birth time. This does not by itself force the model's output, but it makes the evaluation target an annotation of the input stream, so the claimed second-level precision is not independently grounded.

full rationale

The clip-classification experiment (Exp. 1) is a standard supervised pipeline with external backbone architectures, a train/validation/test split, and reported precision/recall/MCC; no circularity is evident there. The circularity concern is concentrated in Exp. 2. First, the ToB estimator is Eq. 4, whose threshold gamma is selected from the FPR of the same test videos (Fig. 4a), and the resulting Table 2 numbers are then reported as system performance; this is a test-set-fitted threshold presented as a prediction. Second, the ground-truth timestamps against which the error is measured were created by manually inspecting the thermal videos (Section 2), so the 1 s/1.5 s deviation measures model-human agreement on the same thermal modality. The paper's self-citations to prior NewbornTime work ([10], [11], [15]) are routine project references, not load-bearing external uniqueness theorems, and the GMM normalization is described with enough independent detail. Note also an internal inconsistency: the abstract reports a 1 s median absolute deviation while Table 2 reports Q2 = 1.5 s, which reinforces that the headline numeric is not stable. Overall, the central Exp. 2 claim is partially circular because its headline error and success rate depend on a threshold fitted to the test set and on same-modality annotations; Exp. 1 retains independent content, so the score is 6 rather than higher.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central quantitative claims rest on a small set of hand-chosen or test-tuned parameters (threshold, filter size, window, stride, backbone) and on several domain assumptions about thermal video as a reliable and privacy-preserving source of ground truth. No new physical entities are introduced.

free parameters (5)
  • Confidence threshold gamma = 0.9
    Eq. 4; chosen by inspecting FPR-vs-threshold curve on the 25-video test set (Section 5, Figure 4a), then used to report final ToB performance.
  • FIR filter size K = 3
    Section 4.2, Eq. 3; hand-chosen smoothing window over prediction scores.
  • Inference window F and stride tau = F=25 frames (3 s), tau=1 s
    Section 4.2; chosen to meet model input and second-precision requirement; affects temporal resolution of ToB estimate.
  • GMM clipping range empirical constraints = not fully specified
    Section 3.1; temperature range of interest is defined based on observations from the delivery room; constraints are not stated numerically.
  • MoViNet-A2 backbone selection = MoViNet-A2
    Selected as best among I3D, X3D, MoViNet variants on the validation set (Table 1); model choice is data-driven.
assumptions (4)
  • domain assumption Manual annotation of ToB from thermal video constitutes ground truth
    Section 2: manual annotation of the ToB with second precision was done by carefully inspecting the thermal videos.
  • domain assumption Birth is visually identifiable in ceiling-mounted thermal video through skin-temperature contrast and motion
    Section 1 and Figure 2; the entire detection approach depends on this.
  • domain assumption The 25 manually selected test videos are representative of the birth population
    Section 4: the test set consists of 25 manually selected thermal videos, keeping the same distribution of the maternal position.
  • standard math GMM with three components separates background, clothing, and skin in thermal videos
    Section 3.1; a modeling assumption about pixel temperature distributions, not proven for all videos.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AI-Based Thermal Video Analysis in Privacy-Preserving Healthcare: A Case Study on Detecting Time of Birth." pith.science (2026). https://pith.science/paper/RUUIZWFV

@misc{pith2026250204365,
  author       = {Pith},
  title        = {Pith review of: AI-Based Thermal Video Analysis in Privacy-Preserving Healthcare: A Case Study on Detecting Time of Birth},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RUUIZWFV}},
  note         = {Machine review of arXiv:2502.04365}
}
read the original abstract

Approximately 10% of newborns need some assistance to start breathing and 5\% proper ventilation. It is crucial that interventions are initiated as soon as possible after birth. Accurate documentation of Time of Birth (ToB) is thereby essential for documenting and improving newborn resuscitation performance. However, current clinical practices rely on manual recording of ToB, typically with minute precision. In this study, we present an AI-driven, video-based system for automated ToB detection using thermal imaging, designed to preserve the privacy of healthcare providers and mothers by avoiding the use of identifiable visual data. Our approach achieves 91.4% precision and 97.4% recall in detecting ToB within thermal video clips during performance evaluation. Additionally, our system successfully identifies ToB in 96% of test cases with an absolute median deviation of 1 second compared to manual annotations. This method offers a reliable solution for improving ToB documentation and enhancing newborn resuscitation outcomes.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

  1. [1]

    Hege Langli Ersdal, Estomih Mduma, Erling Svensen, and Jef- frey M Perlman, “Early initiation of basic resuscitation in- terventions including face mask ventilation may reduce birth asphyxia related mortality in low-income countries: a prospec- tive descriptive observational study,” Resuscitation, vol. 83, no. 7, pp. 869–873, 2012

  2. [2]

    Perinatal asphyxia,

    World Health Organization, “Perinatal asphyxia,” https://www.who.int/teams/maternal-newborn-child- adolescent-health-and-ageing/newborn-health/perinatal- asphyxia, [Online; accessed 11-September-2024]

  3. [3]

    The first golden minute—is it relevant?,

    Tonia Branche, Marta Perez, and Ola D Saugstad, “The first golden minute—is it relevant?,” Resuscitation, vol. 156, pp. 284–285, 2020

  4. [4]

    European resuscitation council guide- lines 2021: Newborn resuscitation and support of transition of infants at birth,

    John Madar, Charles C Roehr, Sean Ainsworth, Hege Ersdal, Colin Morley, et al., “European resuscitation council guide- lines 2021: Newborn resuscitation and support of transition of infants at birth,” Resuscitation, vol. 161, pp. 291–326, 2021

  5. [5]

    Myra H Wyckoff, Robert Greif, Peter T Morley, Kee-Chong Ng, Theresa M Olasveengen, et al., “2022 international consensus on cardiopulmonary resuscitation and emergency cardiovascular care science with treatment recommendations: Summary from the basic life support; advanced life support; pediatric life support; neonatal life support; education, imple- ment...

  6. [6]

    Timing of interventions in the delivery room: Does reality compare with neonatal resus- citation guidelines?,

    Lisa K. McCarthy, Colin J. Morley, Peter G. Davis, C. Omar F. Kamlin, and Colm P.F. O’Donnell, “Timing of interventions in the delivery room: Does reality compare with neonatal resus- citation guidelines?,” The Journal of Pediatrics, vol. 163, no. 6, pp. 1553–1557.e1, 2013

  7. [7]

    Activity recog- nition from newborn resuscitation videos,

    Øyvind Meinich-Bache, Simon Lennart Austnes, Kjersti En- gan, Ivar Austvoll, Trygve Eftestøl, Helge Myklebust, Simeon Kusulla, Hussein Kidanto, and Hege Ersdal, “Activity recog- nition from newborn resuscitation videos,” IEEE journal of biomedical and health informatics , vol. 24, no. 11, pp. 3258– 3267, 2020

  8. [8]

    Comparative analysis of binary and multiclass activity recog- nition in high-quality newborn resuscitation videos,

    Jorge Garc ´ıa-Torres, Øyvind Meinich-Bache, Siren Irene Ret- tedal, Amalie Kibsgaard, Sara Brunner, and Kjersti Engan, “Comparative analysis of binary and multiclass activity recog- nition in high-quality newborn resuscitation videos,” in North- ern Lights Deep Learning Conference 2024, 2023

Show all 21 references
  1. [9]

    Newborn time-improved newborn care based on video and artificial intelligence-study protocol,

    Kjersti Engan, Øyvind Meinich-Bache, Sara Brunner, Helge Myklebust, Chunming Rong, Jorge Garc´ıa-Torres, Hege L Ers- dal, Anders Johannessen, Hanne Markhus Pike, and Siren Rettedal, “Newborn time-improved newborn care based on video and artificial intelligence-study protocol,”...

  2. [10]

    Ad- vancing newborn care: Precise birth time detection using ai- driven thermal imaging with adaptive normalization,

    Jorge Garc ´ıa-Torres, Øyvind Meinich-Bache, Anders Johan- nessen, Siren Rettedal, Vilde Kolstad, and Kjersti Engan, “Ad- vancing newborn care: Precise birth time detection using ai- driven thermal imaging with adaptive normalization,” 2024

  3. [11]

    Consent-driven, semi-automated data collec- tion during birth and newborn resuscitation: Insights from the NewbornTime study,

    Sara Brunner, Anders Johannessen, Jorge Garc ´ıa-Torres, Fer- hat ¨Ozgur Catak, Øyvind Meinich-Bache, Siren Rettedal, and Kjersti Engan, “Consent-driven, semi-automated data collec- tion during birth and newborn resuscitation: Insights from the NewbornTime study,” jan 2025

  4. [12]

    MOBOTIX, Technical Specifications MOBOTIX S16B Du- alFlex, MOBOTIX AG, 12 2021

  5. [13]

    Digital health innovations, tools, and resources to support helping babies sur- vive programs,

    Sherri L Bucher, Peter Cardellichio, Naomi Muinga, Jackie K Patterson, Anu Thukral, Ashok K Deorari, Santorino Data, Rachel Umoren, and Saptarshi Purkayastha, “Digital health innovations, tools, and resources to support helping babies sur- vive programs,” Pediatrics, vol. 146,...

  6. [14]

    Detection of time of birth and cord clamping using thermal video in the delivery room,

    Vilde Kolstad, Jorge Garc ´ıa-Torres, Sara Brunner, Anders Jo- hannessen, Elizabeth Foglia, Hege Ersdal, Øyvind Meinich- Bache, and Siren Rettedal, “Detection of time of birth and cord clamping using thermal video in the delivery room,” Frontiers in Pediatrics, vol. 12, pp. 13...

  7. [15]

    To- wards using thermal cameras in birth detection,

    Jorge Garc ´ıa-Torres, Øyvind Meinich-Bache, Sara Brunner, Anders Johannessen, Siren Rettedal, and Kjersti Engan, “To- wards using thermal cameras in birth detection,” in 2022 IEEE 14th Image, Video, and Multidimensional Signal Processing Workshop (IVMSP). IEEE, 2022, pp. 1–5

  8. [16]

    Quo vadis, action recognition? a new model and the kinetics dataset,

    Joao Carreira and Andrew Zisserman, “Quo vadis, action recognition? a new model and the kinetics dataset,” inproceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 6299–6308

  9. [17]

    X3d: Expanding architectures for efficient video recognition,

    Christoph Feichtenhofer, “X3d: Expanding architectures for efficient video recognition,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 203–213

  10. [18]

    Movinets: Mobile video networks for efficient video recognition,

    Dan Kondratyuk, Liangzhe Yuan, Yandong Li, Li Zhang, Mingxing Tan, Matthew Brown, and Boqing Gong, “Movinets: Mobile video networks for efficient video recognition,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 16020–16030

  11. [19]

    4, Springer, 2006

    Christopher M Bishop and Nasser M Nasrabadi, Pattern recog- nition and machine learning, vol. 4, Springer, 2006

  12. [20]

    Optimal thresholding of classifiers to maximize f1 measure,

    Zachary C Lipton, Charles Elkan, and Balakrishnan Naryanaswamy, “Optimal thresholding of classifiers to maximize f1 measure,” in Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2014, Nancy, France, September 15-19, 2014. Proceedings, Part ...

  13. [21]

    The advantages of the matthews correlation coefficient (mcc) over f1 score and accu- racy in binary classification evaluation,

    Davide Chicco and Giuseppe Jurman, “The advantages of the matthews correlation coefficient (mcc) over f1 score and accu- racy in binary classification evaluation,” BMC genomics, vol. 21, pp. 1–13, 2020

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.